A personal library of Claude Code skills, sub-agents, and a status line for daily engineering, coursework, and client work.
Claude Toolkit is my personal library of extensions for Claude Code — the skills, sub-agents, and status line I actually run day to day across software projects, university coursework, and client web work. It's less a single product than a snapshot of how I've shaped the tool around my own workflow, published so the patterns are there to borrow from.
What's inside
- 17 skills — self-contained folders Claude loads on demand when a request matches their trigger description, covering everything from branching and PR review to Word document editing to grade math from a syllabus.
- 8 sub-agents — specialized assistants with their own model and tool access, for code review, frontend and full-stack work, refactoring, and performance profiling.
- A status line — an Agnoster-inspired bash script showing user, host, git branch, model, context usage, and rate-limit percentage.
The eval harness
The most substantial piece is skill-creator, a meta-skill for building and grading other skills. It runs a real evaluation loop: it spins up throwaway claude -p sessions with a candidate skill installed, fires a battery of trigger queries at each one, and scores whether the skill's description actually caused Claude to read it — a measured trigger rate, not a guess.
# run_eval.py — one query, one measured outcome
def run_single_query(query, skill_name, skill_description, timeout, project_root, model=None):
...
Results roll up through aggregate_benchmark.py into pass/fail summaries across a whole test matrix, and grade_structure.py runs the deterministic parts of a structural rubric — line limits, naming, frontmatter shape — before a grading subagent handles the subjective parts.
Other highlights
- The
docxskill's OOXML validators walk a Word document's XML against the real ECMA-376/ISO schemas, catching broken relationship IDs and duplicate content before a document ships. sapling-ai-detectorcalls a third-party AI-detection API and turns the response into a per-sentence markdown report, with cost tracking built in.- Every skill and sub-agent ships as plain markdown plus scripts — no framework, just files Claude reads when it needs them.